JP2018022119A - Sound source separation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound source separation device capable of reducing a delay time in performing sound source separation on-line while maintaining separation and extraction performance of a sound source.SOLUTION: A sound source separation device 1 includes: a separation matrix processing part 20 for converting observation signals x1(n), x2(n) output from microphones 10, 11 for collecting sounds propagated from a plurality of sound sources into a signal group of a frequency region, updating a separation matrix W(ω) on the basis of the signal group, and converting the updated separation matrix into a time series of filter coefficients ha(n) to output the filter coefficients ha(n); a filter coefficient conversion part 30 for partially removing a non-causal component included in the filter coefficient ha(n) to perform conversion into a filter coefficient h(n); and a separation part 40 for supplying the filter coefficient h(n) to filter groups 41 to 44, and generating a plurality of separation signals y1(n), y2(n) separated from the observation signals x1(n), x2(n) in correspondence with the separation matrix W(ω).SELECTED DRAWING: Figure 1

Description

  The present invention relates to a sound source separation device that separates or extracts the sound of each sound source from a mixed sound of a plurality of sound sources.

  In general, a sound source separation technique for separating or extracting only sound that arrives from a target sound source in a space where various sound sources exist is known. In the case of general sound source separation technology, the target sound source is obtained by storing multiple observation signals obtained by collecting the mixed sounds of multiple sound sources using multiple microphones installed in the space, and performing arithmetic processing. A separation method is adopted. For example, a technique for stably obtaining a separation matrix for separating a target sound source offline based on independent vector analysis has been proposed for a plurality of observation signals (see, for example, Patent Document 1). In addition, for example, a technique for separating a target sound source online by estimating an auxiliary variable for updating a separation matrix from past observation signals has been proposed (see, for example, Patent Document 2).

JP2013-68938A JP 2014-41308 A

  The conventional sound source separation technique has various uses. For example, a sound source separation technique applicable to a general hearing aid is desired. When using a hearing aid, there is a problem of time delay from when an external sound reaches the microphone until the sound is output from the earphone in the auricle. For example, in order for a user to comfortably use a hearing aid, the above-mentioned time delay tolerance limit is considered to be approximately 10 ms. However, as described above, since the technique of Patent Document 1 performs off-line calculation processing, reduction of the delay time from the input of the observation signal to the output of the separated signal is not considered. The technique of the above-mentioned Patent Document 2 can perform arithmetic processing in real time online, but the Fourier transform process and the inverse Fourier transform process are interposed in the main path of the observation signal, so that the total is several hundred ms. Since a delay time of a certain degree occurs, it is difficult to apply to a hearing aid.

  The present invention has been made to solve these problems, and when separating mixed sound of a plurality of sound sources including a target sound source, delay time when performing sound source separation online while maintaining sound source separation performance An object of the present invention is to provide a sound source separation device that can reduce noise.

  In order to solve the above problems, a sound source separation device (1) according to the present invention is a sound source separation device that separates sounds of each sound source from a mixed sound of a plurality of sound sources, and collects sound propagated from the plurality of sound sources. Then, a plurality of microphones (10, 11) that respectively convert into electric signals and a plurality of observation signals (x1 (n), x2 (n)) output in time series from the plurality of microphones are signal groups in the frequency domain. (X1 (ω), x2 (ω)), the separation matrix (W (ω)) for separating the plurality of sound sources is updated based on the signal group in the frequency domain, and the updated separation A separation matrix processing unit (20) for converting a matrix into a time-series filter coefficient (ha (n)) and outputting the matrix, and by partially removing non-causal components included in the time-series filter coefficient, Filter coefficient change to convert filter coefficient The filter coefficient (h (n)) converted by the unit (30) and the filter coefficient conversion unit is supplied to a filter group (41 to 44) that performs a convolution operation of the plurality of observation signals, and the separation matrix Correspondingly, a separation unit (40) that generates a plurality of separated signals (y1 (n), y2 (n)) separated from the plurality of observation signals is configured.

  According to the sound source separation device of the present invention, a plurality of observation signals obtained via a plurality of microphones are branched to the side branch side, and the separation matrix is updated in the frequency domain, and the updated separation matrix is also obtained. Corresponding time domain filter coefficients are supplied to the filter group of the separation unit after the non-causal components are partially removed. Therefore, the separation unit located in the main path of multiple observation signals does not directly update the separation matrix and uses a shortened filter coefficient, so the signal propagation delay time is greatly increased while maintaining the separation performance. It is possible to perform an online operation while reducing the number of times.

  In the present invention, it is possible to provide a plurality of receivers that respectively convert the plurality of separated signals generated by the separation unit into sound. For example, assuming a general hearing aid, the present invention can be applied to a hearing aid including a receiver that outputs sound to the user's external auditory canal.

  The separation matrix processing unit of the present invention includes a short-time Fourier transform unit that converts each of the plurality of observation signals into the frequency domain signal group, and a separation matrix that updates the separation matrix based on the frequency domain signal group. An updating unit and an inverse Fourier transform unit that converts the separation matrix into the time-series filter coefficients can be included. Therefore, since the separation matrix is updated in the frequency domain from the short-time Fourier transform unit to the inverse Fourier transform unit, it is possible to realize arithmetic processing with little influence on the delay time.

  The filter coefficient conversion unit according to the present invention removes a predetermined portion from the cyclic shift unit that performs a cyclic shift on the filter coefficient generated by the separation matrix processing unit, and the non-causal component of the cyclically shifted filter coefficient. And a filter coefficient removing unit to be configured. In this case, it is preferable that the filter coefficient removing unit removes the predetermined portion excluding the predetermined number of samples in the vicinity of the center of the time series among the non-causal components. Therefore, the filter coefficient can be shortened by selectively removing a predetermined portion of the non-causal component that contributes relatively little to the separation performance, and the delay time in the separation unit can be reliably reduced.

  The predetermined number of samples excluded from the predetermined part of the non-causal component is preferably set to the number of samples corresponding to a time not exceeding 10 ms. This is to ensure the performance when the present invention is applied to a hearing aid because the upper limit of the allowable delay time in a general hearing aid is said to be about 10 ms.

  The separation unit of the present invention includes a plurality of FIR filter units corresponding to elements of the separation matrix and a plurality of addition units that add outputs of the plurality of FIR filter units in response to addition of separation operations by the separation matrix. And can be configured. Therefore, a time domain separation unit can be configured by providing K × M FIR filter units and M addition units corresponding to the K × M separation matrix. In the case of a 2 × 2 separation matrix, the separation unit may be provided with four FIR filter units and two addition units (see FIG. 1).

  As described above, according to the present invention, the frequency domain separation operation is performed on the side branch side, and the non-causal components of the time domain filter coefficients are partially removed and then supplied to the filter group of the separation unit. As a result, the signal propagation delay time can be significantly reduced without degrading the separation performance. Therefore, sound source separation can be performed online in real time, and a sound source separation device suitable for application to a hearing aid or the like can be realized.

It is a block diagram which shows the example of a schematic structure of the sound source separation apparatus which is one Embodiment to which this invention is applied. It is a figure which shows the example of a waveform of the filter coefficient output from an inverse Fourier-transform part. It is a figure which shows the example of a waveform of the filter coefficient output from a cyclic shift part. It is a figure which shows the example of a waveform of the filter coefficient after removing by the filter coefficient removal part. It is a comparative example which shows the verification result of the simulation by the sound source separator which applied the conventional structure for contrast with this invention. It is an Example which shows the verification result of the simulation similar to FIG. 5 by the sound source separation apparatus of this embodiment. It is a figure which shows the outline | summary of the conventional structure used in the comparative example of FIG. It is a modification of the sound source separation apparatus of this embodiment.

  Hereinafter, embodiments of a sound source separation apparatus to which the present invention is applied will be described with reference to the accompanying drawings. However, the embodiments described below are examples of forms to which the technical idea of the present invention is applied, and the present invention is not limited by the contents of the present embodiments.

  FIG. 1 is a block diagram illustrating a schematic configuration example of a sound source separation device 1 according to an embodiment to which the present invention is applied. The sound source separation device 1 of FIG. 1 includes two microphones 10 and 11, two receivers 12 and 13, a short-time Fourier transform unit 21, a separation matrix update unit 22, an inverse Fourier transform unit 23, and a cyclic shift unit. 31, a filter coefficient removing unit 32, four FIR filter units 41, 42, 43 and 44, and two adding units 45 and 46. Among these, the short-time Fourier transform unit 21, the separation matrix update unit 22, and the inverse Fourier transform unit 23 constitute the separation matrix processing unit 20, and the cyclic shift unit 31 and the filter coefficient removal unit 32 comprise the filter coefficient conversion unit 30. The FIR filter units 41, 42, 43, 44 and the addition units 45, 46 constitute a separation unit 40.

  In the above configuration, the microphones 10 and 11 are respectively arranged at two different observation positions in a space where a plurality of sound sources exist, collect input sound and convert it into an electric signal. In FIG. 1, one microphone 10 outputs a time-series observation signal x1 (n), and the other microphone 11 outputs a time-series observation signal x2 (n). Separation signals y1 (n) and y2 (), which are signals obtained by estimating the sound of each sound source, with respect to the observation signals x1 (n) and x2 (n) output from the microphones 10 and 11, via a separation unit 40 described later. n) is separated. One receiver 12 converts the separated signal y1 (n) into sound and outputs it, and the other receiver 13 converts the separated signal y2 (n) into sound and outputs it. The receivers 12 and 13 are composed of one receiver and a changeover switch (not shown), and the separated signals y1 (n) and y2 (n) are switched by a switch or the like and output to the receiver. The separated signals y1 (n) and y2 (n) may be converted into sound and output. Furthermore, when the separated signals y1 (n) and y2 (n) are directly taken in and processed by another device, a configuration in which the pair of receivers 12 and 13 in FIG. 1 is omitted may be employed.

  Here, as shown in FIG. 1, only the separation unit 40 is arranged in the main path from the microphones 10 and 11 to the receivers 12 and 13, and the separation matrix processing unit 20 is arranged in a path (side branch) parallel thereto. And the filter coefficient conversion part 30 is arrange | positioned. With such a configuration, until the separated signals y1 (n) and y2 (n) are obtained based on the observation signals x1 (n) and x2 (n), the time delay due to the processing in the side branch described above is affected. Therefore, the delay time can be shortened. The delay time of the sound source separation device 1 of the present embodiment will be described in detail later.

  As shown in FIG. 1, the observation signals x1 (n) and x2 (n) output from the microphones 10 and 11 branch to the side branch side and are input to the short-time Fourier transform unit 21. The short-time Fourier transform unit 21 performs a short-time Fourier transform process by multiplying a predetermined window function while shifting the time-series observation signals x1 (n) and x2 (n), and the frequency-domain observation signal x1 (ω). , X2 (ω). In the short-time Fourier transform unit 21, a predetermined number of consecutive samples of the observation signals x1 (n) and x2 (n) are collectively subjected to transform processing. Observation signals x1 (ω) and x2 (ω) generated by the short-time Fourier transform unit 21 are input to the separation matrix update unit 22.

ただし、Ｔは行列の転置を表す。 The separation matrix updating unit 22 updates the separation matrix W (ω) for generating the separation signals y1 (ω) and y2 (ω) based on the observation signals x1 (ω) and x2 (ω). In the example of FIG. 1, the separation matrix W (ω) is represented by a 2 × 2 matrix, and the following (1) between the observed signal vector x (ω) and the separated signal vector y (ω) in the frequency domain. The relationship of the formula holds.

However, T represents transposition of a matrix.

  Although various separation algorithms can be applied to the separation matrix update unit 22, for example, a well-known independent vector analysis can be applied. This independent vector analysis is a method of updating the separation matrix so that the separation signal vectors in the frequency domain are statistically independent from each other. Independent vector analysis, in principle, has the advantage of avoiding the permutation problem of rearranging the separated signals in each frequency band in association with each sound source.

ただし、
Ｗ：分離行列Ｗ（ω）の全周波数が集合した分離行列
Ｅ［・］：時刻ｔに関する期待値
Ｇ［・］：Ｇ（ｙ（ω））＝−logｑ（ｙ（ω））（音源の確率密度関数ｑ［・］を用いた関数）
ｙ_ｋ：全周波数の分離信号ベクトル
Ｎ_ω：周波数の上限 Here, assuming that independent vector analysis is applied in a situation where there are K sound sources and M observation points, the number of elements of the observation signal vector x (ω) is M and the separated signal vector y (ω). The number of elements is K, and the K × M separation matrix W (ω) is updated. In this case, FIG. 1 is a configuration example corresponding to the case where K = M = 2. The separation algorithm based on independent vector analysis results in a process for minimizing the objective function J (W) shown in the following equation (2).

However,
W: Separation matrix in which all frequencies of the separation matrix W (ω) are aggregated E [•]: Expected value for time t G [•]: G (y (ω)) = − logq (y (ω)) (sound source Function using probability density function q [•]
y _k : separation signal vector of all frequencies N _ω : upper limit of frequency

In order to minimize the objective function J (W) in the equation (2), a predetermined step size η is set, and the correction amount ΔW obtained by the calculation is used to sequentially update W according to the following equation (3). Can be performed.
W ← W−ηΔW (3)
There are various arithmetic algorithms for minimizing the objective function J (W) in the equation (2). For example, an auxiliary function method can be used from the viewpoint of improving the convergence speed. The auxiliary function method is a method for obtaining a separation matrix W for reducing the objective function J (W) by minimizing the auxiliary function set for the objective function J (W), and has a fast convergence. Is a feature.

  Next, returning to FIG. 1, the separation matrix W (ω) updated by the separation matrix update unit 22 is input to the inverse Fourier transform unit 23. The inverse Fourier transform unit 23 performs an inverse Fourier transform process on each element of the separation matrix W (ω) to generate a time domain filter coefficient ha (n). Specifically, based on W11 (ω), W12 (ω), W21 (ω), and W22 (ω), which are elements of a 2 × 2 separation matrix W (ω), four time domain filter coefficients ha11 ( n), ha12 (n), ha21 (n), ha22 (n) are generated. In the following description, when the filter coefficient ha (n) is simply expressed, it is assumed that each of the above-described four filter coefficients ha11 (n), ha12 (n), ha21 (n), ha22 (n) is representative. To do.

Next, the filter coefficient ha (n) generated by the inverse Fourier transform unit 23 is input to the cyclic shift unit 31. The cyclic shift unit 31 performs a so-called cyclic shift on the filter coefficient ha (n), and outputs the filter coefficient hb (n) shifted in the time domain. More specifically, the time series filter coefficients ha (1) to ha (N) corresponding to N samples (n = 1 to N) are sequentially converted by the following equation (4).

  Here, FIG. 2 shows a waveform example of the filter coefficient ha (n) output from the inverse Fourier transform unit 23, and FIG. 3 shows the filter coefficient hb (n) output from the cyclic shift unit 31. An example of a waveform is shown. 2 and 3, the total number of samples on the horizontal axis is N = 4096, and the level on the vertical axis is normalized by −1 to 1. In FIG. 2, n = 0 is shown, but no data actually exists in this portion, and the same applies to the following drawings. First, the filter coefficient ha (n) in FIG. 2 is similar to the waveform of the impulse signal, and many components are concentrated in the vicinity of n = 1 and n = N. Then, when the filter coefficient ha (n) is cyclically shifted by the cyclic shift unit 31, the right half part of FIG. 2 moves to the left part of n = 1 in FIG. As a result, as shown in FIG. 3, in the case of the converted filter coefficient hb (n), a symmetric waveform in which components are concentrated at the approximate center in the range of n = 1 to N is obtained. This cyclic shift is a process to be executed prior to performing a time domain convolution operation in the separation unit 40 described later.

  Next, the filter coefficient hb (n) after the cyclic shift by the cyclic shift unit 31 is input to the filter coefficient removal unit 32. The filter coefficient removal unit 32 removes a predetermined part of the non-causal components from the filter coefficient hb (n) after the cyclic shift, and generates a filter coefficient h (n) with a reduced number of samples. Here, the position P is shown in FIG. 3 described above, and the portion on the left side of the position P in FIG. FIG. 4 shows a waveform example of the filter coefficient h (n) after being removed by the filter coefficient removing unit 32, and the vertical axis and the horizontal axis follow the same notations as in FIGS.

  Returning to FIG. 3, when the sample at the center (n = 2048) of the filter coefficient hb (n) is regarded as the sample at time t = 0, the second half of the time domain (range from the center to the right) is the filter operation. Since a past sample on the time series is used, it can be said that it is a range having causality. On the other hand, the first half of the time domain (the range from the center to the left) is a non-causal range because future samples on the time series are used in the filter operation. In the present embodiment, the filter coefficient removal unit 32 removes samples of the input filter coefficient hb (n) excluding the predetermined number of non-causal components, thereby reducing the delay time associated with the arithmetic processing. It was confirmed that it can be reduced.

  Note that removing all of the non-causal components from the filter coefficient hb (n) causes performance degradation of sound source separation, so that only a predetermined number of samples in the vicinity of the center in the time series is left. It is. For example, in the example of FIG. 4, a portion corresponding to 160 samples in the vicinity of the position P in FIG. 3 is left, and the range on the left side is removed. In this case, assuming that the total number of samples is N = 4096, the number of samples to be removed is 4096 / 2-160 = 1888, and as a result, the number of samples of the filter coefficient h (n) is 4096/2 + 160 = 2208.

  Here, when applying the sound source separation device 1 of the present embodiment to a hearing aid, it is desirable to set the predetermined number of samples to be left among the above-mentioned non-causal components to a number of samples corresponding to a time not exceeding 10 ms. When the predetermined number of samples exceeds 10 ms in terms of time, the delay time from the microphones 10 and 11 to the receivers 12 and 13 also exceeds 10 ms. It is known that the allowable delay time in a general hearing aid does not exceed 10 ms. That is, if the predetermined number of samples is set to a time exceeding 10 ms, it may cause a sense of discomfort to the user of the hearing aid to which the sound source separation device 1 of the present embodiment is applied. On the other hand, the lower limit value of the predetermined number of samples to be left among the above-mentioned non-causal components is related to the separation performance, but is appropriately set according to conditions such as the use situation of the hearing aid and the separation performance.

  Next, the converted filter coefficient h (n) obtained by the filter coefficient removal unit 32 is supplied to the FIR filter units 41 to 44 included in the separation unit 40. Specifically, the filter coefficient h11 (n) is supplied to the FIR filter unit 41, the filter coefficient h12 (n) is supplied to the FIR filter unit 42, and the filter coefficient h21 (n) is supplied to the FIR filter unit 43. The filter coefficient h22 (n) is supplied to the FIR filter unit 44. Among them, the time series observation signal x1 (n) is input to the two FIR filter sections 41 and 43, and the time series observation signal x2 (n) is input to the two FIR filter sections 42 and 44. The role of the FIR filter units 41 to 44 is to perform a convolution operation in the time domain corresponding to each product included in the above-described equation (1) in the frequency domain.

  Subsequently, two addition units 45 and 46 are arranged at the subsequent stage of the FIR filter units 41 to 44. One adding unit 45 adds the outputs of the two FIR filter units 41 and 42, and outputs the addition result as a separated signal y1 (n). The other addition unit 46 adds the outputs of the two FIR filter units 43 and 44, and outputs the addition result as a separated signal y2 (n). The roles of the adding units 45 and 46 are to perform addition included in the above-described equation (1) in the frequency domain in the time domain. Of the separated signals y1 (n) and y2 (n) obtained by the FIR filter units 41 to 44 and the adding units 45 and 46, one separated signal y1 (n) is converted into sound via the receiver 12, and the other Separated signal y2 (n) is converted into sound via the receiver 13.

In the separation operation in the separation unit 40, when the observation signals x1 (n) and x2 (n) are input, the separation signals y1 (n) and y2 (n) are output, and the number of taps is T, the following (5 ) And (6).

  The sound source separation device 1 in FIG. 1 has two observation signals x1 (n) and x2 (n), two separation signals y1 (n) and y2 (n), and a 2 × 2 separation matrix W (ω ), But more generally, a sound source based on M observed signals x (n), K separated signals y (n), and a K × M separation matrix W The present invention can also be applied to a sound source separation device that performs separation. In this case, in the configuration example of FIG. 1, it is necessary to provide M microphones and K receivers, and the separation unit 40 must be provided with K × M FIR filter units and K addition units.

  Next, the effects of the sound source separation device 1 of the present embodiment will be described with reference to FIGS. FIG. 5 is a comparative example showing a simulation verification result by a sound source separation device to which a conventional configuration is applied for comparison with the present invention, and FIG. 6 shows a similar simulation by the sound source separation device 1 of the present embodiment. It is an Example which shows a verification result. Here, FIG. 7 shows an outline of a conventional configuration used in the comparative example of FIG. That is, two microphones 10 and 11 and two receivers 12 and 13 similar to the configuration in FIG. 1 are provided, and the short-time Fourier transform unit 100, the separation matrix calculation unit 101, and the inverse are sequentially arranged in the main path between them. A short-time Fourier transform unit 102 is arranged.

  In the simulations of FIGS. 5 and 6, a mixed sound of male voice and female voice is input to the microphones 10 and 11, and separation operation based on the observation signal x 1 (n) of the microphone 10 and the observation signal x 2 (n) of the microphone 11 is performed. The result was generated by generating two separated signals y1 (n) and y2 (n). 5 and 6 show the waveforms of the observation signal x1 (n) and the separation signals y1 (n) and y2 (n) within the time range of 0 to 0.7 s. 5 and 6 do not show the waveform of the original signal before mixing, but the waveforms of the separated signals y1 (n) and y2 (n) are almost faithful to the original signal and mainly have a delay time. Only matters.

  First, in the conventional comparative example of FIG. 5, the delay time of the separated signals y1 (n) and y2 (n) with respect to the observation signal x1 (n) is about 0.3 s (300 ms). On the other hand, according to FIG. 6 of the present embodiment, the delay time of the separated signals y1 (n) and y2 (n) with respect to the observation signal x1 (n) is extremely small, and is approximately 0.01 s (10 ms). It was confirmed. Such a significant reduction in the delay time is that the delay of the main path is eliminated because the separation matrix W is updated on the side branch side, and in addition, the filter coefficient h (n) is filtered by the filter coefficient conversion unit 30. This is because non-causal components that contribute relatively little to sound source separation are removed. Therefore, by adopting the configuration of the sound source separation device 1 of the present embodiment, it is possible to solve the problems associated with the time delay between input and output while performing online arithmetic processing. For example, the sound source separation device 1 suitable for application to a hearing aid Can be realized.

  Next, FIG. 8 shows a modification of the sound source separation device 1 of the present embodiment. The modification in FIG. 8 assumes a so-called projection back configuration, and not only separates the mixed sound collected by the two microphones 10 and 11, but also holds the localization information of the sound source like a stereo device. It has a function of separating as it is. In the modification of FIG. 8, the difference from FIG. 1 is that a projection back calculation unit 50 that performs a known calculation for applying projection back to the separation matrix W (ω) is added to the separation matrix processing unit 20. The point and the path of the separation unit 40 and the pair of separated signals y1 (n) and y2 (n) are two systems.

  That is, the two separation units 40p and 40q both receive two observation signals x1 (n) and x2 (n) from two microphones, and the filter coefficient h (n) corresponding to each from the filter coefficient conversion unit 30. Is supplied. One separating unit 40p outputs two separated signals y1p (n) and y2p (n), and the other separating unit 40q outputs two separated signals y1q (n) and y2q (n). Among these, the localization of one sound source is obtained by a pair of separated signals y1p (n) and y1q (n), and the localization of the other sound source is obtained by a pair of separated signals y2p (n) and y2q (n). .

  The sound source separation device 1 to which the present invention is applied has been described above according to the present embodiment, but the present invention can be applied to various devices. That is, the sound source separation device 1 of the present invention can be applied to a general hearing aid as described above, but may be incorporated as a part of other computers or communication devices. In addition, the configuration of FIG. 1 of the present embodiment can be appropriately changed as long as it has the same function, and in addition to the change of detailed processing content, a network or wireless communication may be interposed in the interconnection between the members. Other points are not limited to the contents of the present embodiment, and various configurations and processes can be employed.

DESCRIPTION OF SYMBOLS 1 ... Sound source separation apparatus 10, 11 ... Microphone 12, 13 ... Receiver 20 ... Separation matrix processing part 21 ... Short-time Fourier transformation part 22 ... Separation matrix update part 23 ... Inverse Fourier transformation part 30 ... Filter coefficient transformation part 31 ... Cyclic shift Unit 32 ... Filter coefficient removing unit 40 ... Separating units 41, 42, 43, 44 ... FIR filter units 45, 46 ... Adding unit

Claims (7)

A sound source separation device for separating the sound of each sound source from the mixed sound of a plurality of sound sources,
A plurality of microphones for collecting sound propagating from the plurality of sound sources and converting the sound into electric signals;
A plurality of observation signals output in time series from the plurality of microphones are converted into frequency domain signal groups, and a separation matrix for separating the plurality of sound sources is updated based on the frequency domain signal groups, and updated. A separation matrix processing unit that converts the separation matrix into a time-series filter coefficient and outputs the filter matrix;
A filter coefficient conversion unit that converts the filter coefficient by partially removing non-causal components included in the time-series filter coefficient;
The filter coefficients converted by the filter coefficient conversion unit are supplied to a filter group that performs a convolution operation of the plurality of observation signals, and a plurality of separated signals separated from the plurality of observation signals corresponding to the separation matrix A separation unit for generating
A sound source separation device comprising:   The sound source separation device according to claim 1, further comprising a plurality of receivers that respectively convert the plurality of separation signals generated by the separation unit into sound. The separation matrix processing unit
A short-time Fourier transform unit that transforms each of the plurality of observation signals into a signal group in the frequency domain;
A separation matrix updating unit that updates the separation matrix based on the signal group in the frequency domain;
An inverse Fourier transform unit for transforming the separation matrix into the time-series filter coefficients;
The sound source separation device according to claim 1, comprising: The filter coefficient conversion unit
A cyclic shift unit that performs a cyclic shift on the filter coefficient generated by the separation matrix processing unit;
A filter coefficient removing unit that removes a predetermined portion of the non-causal components of the cyclically shifted filter coefficients;
The sound source separation device according to claim 1, comprising:   5. The sound source separation device according to claim 4, wherein the filter coefficient removing unit removes the predetermined portion from the non-causal component excluding a predetermined number of samples in the vicinity of the center of the time series.   6. The sound source separation device according to claim 5, wherein the predetermined number of samples is set to a number of samples corresponding to a time not exceeding 10 ms. The separation unit is
A plurality of FIR filter units corresponding to elements of the separation matrix;
In response to the addition of the separation operation by the separation matrix, a plurality of addition units that add the outputs of the plurality of FIR filter units;
The sound source separation device according to claim 1, comprising:

Images